Automatic Positioning of Hippocampus Deformable Mesh Models in Brain MR Images Using a Weighted 3D-SIFT Technique

Abstract

Automatic hippocampus segmentation in Magnetic Resonance (MR) images is an essential step in systems for early diagnostic and monitoring treatment of Alzheimer’s disease (AD). It allows quantification of the hippocampi volume and assessment of their progressive shrinkage, considered as the hallmark symptom of AD. Among several methods published in the literature for hippocampus segmentation, those using anatomical atlases and deformable mesh models are the most promising ones. Although these techniques are convenient ways to embed the shape of the models in the segmentation process, their success greatly depend on the initial positioning of the models. In this work, we propose a new keypoint deformable registration technique that uses a modification of the 3D Scale-Invariant Feature Transform (3D-SIFT) and a keypoint weighting strategy for automatic positioning of hippocampus deformable meshes in brain MR images. Using the Mann-Whitney U test to assess the results statistically, our method showed an average improvement of 11% over the exclusive use of Affine transformation, 30% over the original 3D-SIFT and 7% over the non-weighted point procedure.